A General and Scalable Solution of Heterogeneous Workflow Invocation and Nesting

1. Tamas Kukla, Tamas Kiss, Gabor Terstyanszky Centre for Parallel computing University of Westminster London Peter Kacsuk Computer and Automation Research Institute Hungarian Academy of Sciences Budapest A General and Scalable Solution of Heterogeneous Workflow Invocation and Nesting

2. Contents • Introduction • Approaches to workflow interoperability • Requirements of workflow engine integration • Realising workflow integration • Conclusions

3. Introduction • Several widely utilised, Grid workflow management systems, such as Triana, P-GRADE, Taverna, Kepler, CppWfMS, YAWL, or the K-Wf Grid emerged in the last decade. • These systems were developed by different scientific communities for various purposes. • Therefore, they differ in several aspects. They use • different workflow engines • different workflow description languages • different workflow formalisms • different Grid middleware

4. Different workflow engines • Most systems are coupled with one engine: • Taverna uses Freefluo • Triana uses Triana engine • K-WfGrid uses GWES (Grid workflow execution service)‏ • Older versions of P-GRADE used Condor DAGMan, while its recent version, WS-PGRADE uses its own engine Xen.

5. Different workflow description languages • Most workflow systems use different workflow description languages: • Triana interprets BPEL (Business Process Execution Language) and its own language format. • Taverna workflows are represented in SCUFL. • Older versions of P-GRADE used Condor DAG, recent WS-PGRADE uses its own WS-PGRADE language. • Kepler uses MOML. • YAWL system uses YAWL language. • K-WfGrid uses GWorkflowDL. • Because of this diversity, workflows of a system cannot be reused in another system.

6. Different workflow formalisms • Workflow description languages are based on various workflow formalisms. • Condor DAG uses directed acyclic graphs (DAG)‏. • SCUFL is also DAG based, but it is extended with control constraints. • WS-PGRADE is also DAG based, but it is extended with control constraints, nesting and recursion. • YAWL and GWorkflowDL are based on Petri Nets. • BPEL is Pi-Calculus based. • Different formalisms have different expression capabilities. • Therefore, in many cases it is not possible to express a certain workflow type in the description language of another one.

7. Workflow interoperability • In order to achieve cross-organisational collaboration between the different scientific communities, workflows should be able to interoperate, communicate with and/or invoke each other during execution. • The WfMC (Workflow Management Coalition) defines workflow interoperability in general as: • "The ability for two or more Workflow Engines to communicate and work together to coordinate work."In this definition the workflow engine is a piece of software that provides the workflow run-time environment.

8. Approaches to workflow interoperability • Various solutions can bring workflow interoperability into effect: • Workflow description standardisation • Would enable the exchange of workflows of different systems • XPDL was defined by the WfMC and BPEL was defined by Microsoft and IBM for this purpose, but they did not gain universal acceptance so far. • It is unlikely in the near future • Workflow translation • Would enable the translation from one language to another • Can be realised by translating via an intermediate workflow language. • YAWL and GWorkflowDL could also be used for this purpose. See BPEL to YAWL translator or SCUFL to GWorkflowDL converter. • Cannot be applied in any case

9. Workflow engine integration • An alternative approach to attain workflow interoperability could be realised by workflow engine integration. • Executes the workflow in its native environmentby its own workflow engine. • Makes workflow management systems be able to execute non-native workflows. • Can be realised by loosely or tightly coupled integration.

10. Tightly(i) and loosely(ii) coupled engine integration WF SystemC (i)‏ Engine ofWF System A C C C Engine ofWF System B WF SystemC Engine ofWF System C C I I C I Interface ofWF integrationservice I (ii)‏ Workflow engineintegration service C

11. Workflow engine integration can realise synchronous (i) and asynchronous (ii) workflow execution • (i) - Non-native workflow nesting is a synchronousworkflow execution, where the nested Workflow is executed as a node of the native workflow. • (ii) - Non-native workflow invocation is asynchronous, when the non-native workflow is invokedby a node of the native workflow. Once the execution of the invoked workflow started, there is no further interest in it. Workflow ofsystem A Workflow ofsystem B (i)‏ Workflow ofsystem A Workflow ofsystem B (ii)‏

12. Requirements of workflow engine integration • Our aim is to provide a solution for workflow sharing and interoperability by integrating different workflow systems in the following way: • providing a generic solution, which can be adopted to any workflow system • providing a scalable solution in the sense of both number of workflows and amount of data • integration of a new workflow engine to the system should not require code re-engineering, only user level understanding of the engine in question

13. The concept of the heterogeneous WF system • In a certain type of workflow (home workflow) other types of workflows can be executed as nodes • The home workflow should be any type of WF (Taverna, Triana, Kepler, etc.) • The embedded workflows • should be any kind of WFs • they can work as home workflows for any other type of embedded WFs

14. Realising workflow integration • To provide a generic solution: • It is recommended to realise loosely coupled integration • To provide a scalable solution: • It is recommended to utilize Grid resources for workflow engine execution • To make the workflow engine deployment straightforward: • It is recommended to handle workflow engines as legacy applications

15. Realising workflow integration via a Grid based application repository and submitter • In order to integrate different workflow engines a Grid application repositoryandsubmitterservice, called GEMLCAis used • The reference implementation integrates four different workflow engines (engines of P-GRADE, Taverna, Triana, and Kepler) • Any of these 4 WF systems can be the home WF • Since the integration is based on GEMLCA, the home workflow engine and GEMLCA should be integrated (this is the first step in the integration procedure) • We have already integrated GEMLCA with the P-GRADE workflow system, so P-GRADE was used as the home WF system. • The solution can be adopted by any other workflow system by integrating the GEMLCA web service client to the given system.

16. GEMLCA • GEMLCA is an application repository extended with a job submitter, and allows the deployment of legacy code applications on the Grid. • An application can be exposed via a GEMLCA service and can be executed by using a GEMLCA client. • The legacy application is stored either in the repository of a GEMLCA service or on a third party computational node where GEMLCA can access it. • To publish a legacy application via GEMLCA, only a basic user-level understanding of the legacy application is needed, code re-engineering is not required. • As soon as the application is deployed, GEMLCA is able to submit it using either GT2, GT4 or gLite Grid middle-ware. • If the workflow engine requires credentials to utilise further Grid resources for workflow execution, these are automatically provided by GEMLCA through proxy delegation.

17. Exposing workflow engines via GEMLCA • Command-line workflow engines, just like other legacy applications, can be exposed via a GEMLCA service, without code re-engineering and can be automatically submitted by GEMLCA to the Grid to a computational node. • Three engines (engine of Taverna, Triana, and Kepler) have been installed on our cluster at the University of Westminster on a shared disk so that any cluster node can access them.

18. Realisation of a Workflow engine repository and submitter via GEMLCA Workflow system User selects the required workflow engine, uploads the workflow, the input parameters and input files. The Job manager of the cluster schedules the job to a node. cluster Shared storage WF Engine 1 WF Engine 2 WF Engine 3 GEMLCAclient GEMLCA service Deployed apps Backends WF Engine 1 GT2 WF Engine 2 GT4 WF Engine 3 gLite Executable: WF engine (that is already installed on the cluster)WF to execute: an input parameter of the GEMLCA job

19. Exposing workflow engines via GEMLCA • The engines were en-wrapped by scripts so as to provide a general command line interface for them. This interface is the following:wfsubmit.sh -w wf_descriptor [-p wf_input_params] [-i wf_input_files] [-o wf_output_files]Wrapper scripts are responsible for decompressing the workflow input files, execute the workflow by parametrizing and invoking the workflow engine and finally compress the workflow outputs into one archive file.

20. Parametrization of non-native workflow execution within theP-GRADE portal • GEMLCA was integrated to the P-GRADE portal. • GEMLCA jobs can be parametrized using a JAVA based GUI within the P-GRADE workflow editor. • Any other workflow system can adopt this solution and integrate a GEMLCA client. Selecting Grid Setting workflow descriptor Selecting GEMLCA service Selecting workflow engine Setting input parameters Selecting computational site Setting workflow input files Setting workflow output file

21. Exposing Taverna workflow engine using GEMLCA Administration Portlet • The engines were exposed using the JSR-168 based GEMLCA administrator portlet.

22. Legacy Code interface Description of the exposed Taverna engine

23. Case Study • A case study workflow, that presents how workflows of different systems interoperate, will be presented. • It serves only demonstration purposes, it is not a real life example. • It is a high level heterogeneous P-GRADE workflow, nesting a Taverna, Kepler and Triana workflows. • The data that is transferred between the workflows is stored files, there is no data transformation. • If data transformation is needed, user has to create a data transformer job.

24. Taverna workflow • This workflow fetches several images from a database, creates a few directories and places the images into those directories as image files.

25. Kepler workflow • This workflow goes through the directory structure of the archive input file and manipulates each image that it finds. • The manipulation includes edge highlighting, picture resizing and image type conversion.

26. Triana workflow • This workflow couples the pictures, merges each couple and converts the merged pictures to greyscale images. • Then, one colour component, that can be either the blue, green or red, is taken of the greyscale pictures and saved as new image file.

27. Heterogeneous P-GRADE workflow embedding Triana, Taverna, and Kepler workflows Triana workflow Taverna workflow P-GRADE workflow Kepler workflow

28. Conclusion • This presentation introduced a general solution to workflow interoperability and sharing at the level of workflow integration. • The solution exposes various workflow engines via a GEMLCA service, that is capable of submitting the engines to the Grid. • Hence, it keeps the data at computational sites and offers a solution that is scalable in terms of number of workflows and amount of data. • Workflow engine deployment to this system does not require any code re-engineering, user level understanding is sufficient. • The solution can be adopted by any workflow management system by integrating GEMLCA with the selected WF system